Author
Federico Ulfo
Founder · AI Socratic
Founder of AI Socratic
Website / profile ↗By and about Federico Ulfo · page 3
Tokenmaxxing, 2025-2026, RIP
Meta killed its tokenmaxxing leaderboard as Uber capped AI spending at $1,500/month and GitHub Copilot users saw costs jump 10-50x after switching to per-token billing in June.
newsBun rewrite in Rust
Bun completed a 535,000-line rewrite from Zig to Rust in 11 days using 64 parallel Claude loops, eliminating memory safety bugs that caused crashes when Zig's manual memory management clashed with JavaScriptCore's garbage collector.
newsUnsloth founder's 2h42m fine-tuning masterclass
Unsloth founder, ex-NVIDIA engineer, covers reinforcement learning, kernels, reasoning, quantization, and agents in a 2h42m masterclass.
newsSutton starts Oak Lab
Sutton launches Oak Lab, a new AI company.
newsClaude Sonnet 5: Near-Opus Performance
Anthropic released Claude Sonnet 5 on June 30 with 63.2% SWE-bench Pro and a GDPval-AA score of 1,618—nearly matching Opus 4.8's 1,615—at $2/$10 per million tokens through August 31, with native 1M context and 128K output by default.
news46 thoughts on the near future
AI breakthroughs in autonomous research and robotics will drive deflationary abundance while concentrating power and straining social coordination.
newsFable 5 Is Back: Washington Lifts the 19-Day Ban — With Strings Attached
Commerce Department lifted its 19-day ban on Anthropic's Fable 5 and Mythos 5 on June 30, requiring a >99%-effective safety classifier, pre-release government review, and a cross-lab jailbreak rubric with Amazon, Microsoft, and Google before the models…
newsOpenAI Launches GPT-Live, a New Generation of Voice Models
OpenAI launched GPT-Live, a new generation of voice models for natural human-AI conversation, rolling out in ChatGPT on July 8.
newsOpenAI Floats a 5% Stake for Uncle Sam
OpenAI proposed giving the US government a 5% equity stake worth roughly $42.6B, with a framework for Anthropic, Google, and Meta to cede similar stakes into a sovereign-wealth vehicle, though Anthropic says it hasn't discussed the idea.
newsMeta Ships Muse Spark 1.1
Meta released Muse Spark 1.1 and launched its Meta Model API public preview, claiming the model rivals GPT-5.5 and Opus 4.8 on agentic evals while undercutting on price—Meta's first serious push to monetize its $115–135B infrastructure spend by selling…
newsPrime Intellect raises $130M Series A
Prime Intellect raised $130M Series A led by Radical Ventures to build an open-source stack for training and deploying AI models, with backing from NVIDIA, Intel Capital, and Dell Capital.
newsLiquidAI releases Antidoom, Open-Source Fix for Doom Loops in Reasoning Models
Liquid AI released Antidoom, an open-source method that eliminates "doom loops" in reasoning models by reducing failure rates from 10.2% to 1.4% on LFM2.5-2.6B and 22.9% to 1% on Qwen3.5-4B through targeted retraining of trigger tokens.
newsDario, Fearmongering, and the Real Reason Frontier Models Got Restricted
U.S. government restrictions on frontier models like Fable 5 and GPT-5.6 stem from genuine national-security concerns over cyber capabilities and Chinese access, not primarily Dario Amodei's warnings, though Anthropic's clumsy Pentagon relations made the…
newsUn-0: Coupled Oscillators Replace the Neural Net
Unconventional AI released Un-0, an image generator built on coupled Kuramoto oscillators instead of neural network layers, achieving FID 6.74 on ImageNet 64×64 with 322M parameters—the strongest result yet for a physics-system-based generative model.
newsDSpark: Confidence-Scheduled Speculative Decoding with Semi-Autoregressive Generation
DeepSeek open-sourced DSpark, a confidence-scheduled speculative decoding method that delivers 51% to 400%+ throughput boosts on its V4 models and works on Gemma and Qwen too.
newsOpenAI Launches GPT-5.6 Family: Sol, Terra, and Luna in Limited Preview
OpenAI previewed GPT-5.6 on June 26 — Sol at $5/$30 with a 91.9% Terminal-Bench 2.1 ultra mode, Terra at $2.50/$15 and Luna at $1/$6 — with access limited to ~20 government-vetted organizations.
newsSnap's $2,195 Specs Are Here — and the Glassholes Won
Snap unveiled Specs, a $2,195 standalone AR glasses with a 51° display and four-hour battery, shipping this fall—a decade after we mocked Google Glass.
newsThe Data Black Hole at the Center of AI
Dwarkesh Patel argues AI's real bottleneck is sample efficiency: humans learn from 200 million words while frontier models train on trillions of tokens, a millionfold gap that scaling laws alone may not close.
newsThinking to Recall: How Reasoning Unlocks Parametric Knowledge in LLMs
Google Research finds that enabling reasoning in LLMs helps them recall simple facts through two mechanisms: extra tokens provide computational buffer time, and intermediate facts prime the correct answer—but hallucinated intermediate facts sharply…